Microsoft launched MAI Code 2 and MAI Code 2 Flash at Ignite on November 17, 2026, and the launch post on microsoft.ai leads with a number designed to end conversations: 98% of GPT-6.1 Sol's performance at 40% of the cost, and 96% of Claude Opus 5.5's performance at 5% of the cost. Both claims are, as stated, Microsoft's own measurements. But the interesting fact is not either number by itself. It is what happens when you divide one by the other. MAI Code 2's promotional token price is a quarter of Sol's, yet its cost per task is 40% of Sol's. That gap is arithmetic you cannot argue with: in Microsoft's own evaluation runs, the model consumed roughly 1.6 times as many tokens per task as the model it is undercutting. The discount is real. The token appetite is too.
What Microsoft Actually Shipped
MAI Code 2 is the second generation of Microsoft's in-house coding line, which started with MAI-Code-1-Flash in GitHub Copilot in June 2026. The lineup split is deliberate. MAI Code 2 is the general-purpose agentic coding model, positioned for deep debugging, multi-file implementation, refactoring, and repository-wide reasoning. MAI Code 2 Flash is the lightweight sibling for code explanations, targeted edits, and routine fixes at high volume. The GitHub changelog confirms both are generally available, not previews, in the Copilot model picker across VS Code, Visual Studio, JetBrains, Xcode, Eclipse, the Copilot CLI, the cloud agent, github.com, and GitHub Mobile.
Three details from the launch materials matter for how you should read the claims. First, Microsoft states plainly that these models are "explicitly not yet tuned for harder problems, like fundamental science research and very long horizon agentic work." This is a deliberately scoped model, not a frontier attempt. Second, per the model card coverage by eesel's engineering team, both models are sparse mixture-of-experts architectures with roughly 80B active parameters out of about 2.8T total for MAI Code 2 (12B active of about 398B for Flash), with a 1,050,000-token context window split into 922,000 input and 128,000 output, and five reasoning effort levels. Third, distribution is Copilot-first: MAI Code 2 is not in Microsoft Foundry at all yet, only Flash is in a limited-access preview there, and the weights are not distributed. You cannot plug this model into your own harness today.
The Numbers, With Their Labels Intact
Everything in the table below is Microsoft-reported, at max reasoning effort, from the launch post. DeepSWE 1.1, Terminal-Bench 2.1, and SWE Atlas ran on MSBench, Microsoft's internal evaluation platform, using the GitHub Copilot CLI without subagents. Only FrontierCode v1.1 came from Artificial Analysis, an outside party.
| Benchmark | MAI Code 2 | GPT-6.1 Sol | Claude Opus 5.5 | Who ran it |
|---|---|---|---|---|
| DeepSWE 1.1 | 73.2% | 74.6% | 69.0% | Microsoft |
| Terminal-Bench 2.1 | 84.6% | 85.0% | 78.7% | Microsoft |
| SWE Atlas | 43.1% | 47.4% | 57.5% | Microsoft |
| FrontierCode v1.1 | 52.0% | 49.4% | 56.0% | Artificial Analysis |
| Average pass rate | 63.2% | 64.1% | 65.3% | mixed |
| Avg cost per task | $0.309 (promo) | $0.761 | $6.306 | Microsoft |
The Flash tier tells the same shape of story at lower stakes: average pass rate 61.0% against 55.1% for GPT-6 Luna and 53.9% for Claude Haiku 5.5, at $0.052 per task (promo price) versus $0.080 and $0.319. Pricing in Copilot, per GitHub's pricing docs as summarized by eesel: MAI Code 2 runs $0.50 per million input and $2.50 per million output tokens at promo, and $0.05/$0.25 for Flash, with both prices doubling on May 17, 2027. For comparison, Sol sits at $2/$10 and Opus 5.5 at $4/$20.
One result deserves separate mention because it is harder to game: in blind frontend evaluations across user interfaces, games, data visualizations, 3D scenes, and SVG graphics, annotated by 258 people, reviewers preferred MAI Code 2's output over GPT-6 Sol's 64% of the time, on both aesthetics and functionality.
Our Read: Divide the Two Claims
Here is the analysis most launch coverage skipped, and it is three lines of arithmetic.
Cost per task equals tokens per task times price per token. During the promo, MAI Code 2's token prices are exactly one quarter of Sol's ($0.50 versus $2.00 input, $2.50 versus $10 output). Microsoft's own cost-per-task figures give a ratio of $0.309 divided by $0.761, which is 0.406. So the implied token ratio is 0.406 divided by 0.25, which is about 1.6. In Microsoft's own runs, MAI Code 2 burned roughly 60% more tokens per completed task than Sol. That is not disqualifying, an 80B-active MoE taking more reasoning tokens to land at near-parity is a plausible design trade, but it changes what the discount means. A 75% token-price cut that returns a 59% cost cut is a model efficiency tax of about 16 points, and it is the number your FinOps dashboard will actually show.
The second consequence of that arithmetic is the promo cliff. Every MAI rate doubles on May 17, 2027, while tokens per task do not change. At list price, MAI Code 2's cost per task becomes about $0.618, which is 81% of Sol's $0.761, not 40%. The headline discount has an expiration date baked into the pricing page. Flash is starker: at list price its implied $0.104 per task exceeds Luna's $0.080, so the "65% of the cost" claim flips to roughly 130% of Luna per task once the promo ends, while still scoring six points higher. If you build routing logic on these numbers, the correct default is: MAI Code 2 for volume work now, and a re-benchmark scheduled for May 2027, not a permanent migration.
The third thing a data scientist should flag is where the average hides the distribution. MAI Code 2 beats Opus 5.5 on DeepSWE and Terminal-Bench, then loses SWE Atlas by 14.4 points (43.1% versus 57.5%). If your work is long-horizon, sprawling repo changes, that row is the relevant benchmark and the 96%-at-5% headline is the wrong one to price against. There is also an anomaly worth sitting with: per the launch post, Flash scores 44.8% on SWE Atlas against 43.1% for its bigger sibling. A smaller model beating a larger one on the same benchmark, same harness, same effort level, is either a genuine tuning difference (Flash was optimized for exactly this class of task) or a signal that SWE Atlas scores at this range carry more noise than a 1.7-point gap can survive. Without a third-party rerun, treat that delta as uninformative. And note the comparison set is already stale: Microsoft benchmarked against GPT-6.1 Sol on the same day OpenAI shipped GPT-6.2, which GitHub added to Copilot the same morning.
From the software engineer seat, the strategic read is that this was never primarily an API product. It is a distribution play. Enterprise and Business admins get new models enabled by default under the global model policy unless they have turned that off, which is a change from MAI Code 1.1 Flash, which shipped opt-in in August. Microsoft is routing its in-house model into the default path of the largest installed base of coding-assistant users, then letting price do the switching. That is the same cost-per-task battleground we traced in Kimi K3 1 Pro's 33% price cut and in Meta's Muse Ember launch, where per-token price again turned out to be the least interesting figure once agent loops were counted. The serving economics also fit Microsoft's pattern of squeezing its own silicon and infrastructure: weights stay on Azure, and the company's own MAIA accelerator story, from tokens per megawatt math on MAIA 100, is the back end that makes an 80B-active MoE profitable at $0.50 per million tokens. Whether the in-house research culture can sustain this cadence is an open question we looked at in Ian Langford's departure from Microsoft.
Outlook
The honest summary: MAI Code 2 looks like a genuinely good default for routine, multi-file Copilot work during a six-month promo window, with a real independent-ish signal (the 258-annotator blind frontend preference) behind its quality claims, and three caveats you should carry into any adoption decision. Three of four benchmarks were run by Microsoft on Microsoft's own platform against a previous-generation competitor. The cost advantage compresses from 60% to 19% the day list pricing starts. And the model is unavailable anywhere but Copilot, so the switching decision is bundled with the platform decision.
The signal that would force a rewrite of this piece: an Artificial Analysis page for MAI Code 2 with independently run numbers, or Microsoft publishing tokens-per-task directly instead of leaving the 1.6x ratio as inference. Until then, the 98%-at-40% claim is vendor-reported, internally consistent, and quietly more expensive per token consumed than the headline admits.